device=cuda MIMO: n_ctrl/draw=960 per-ctrl ms per draw=[np.float64(1.1584), np.float64(1.005), np.float64(0.8614)] mean=1.0083 sd=0.1485 pendulum: n_ctrl/draw=320 per-ctrl ms per draw=[np.float64(1.323), np.float64(1.3257), np.float64(1.462)] mean=1.3702 sd=0.0795 === TIMING CONSISTENCY (S5) === Timing consistency audit (S5) -- device=cuda, 3 seeded draws ((11, 12, 13)), >= 960 controllers/draw, batched as the evals batch them (K=8). MIMO 2x2 (models/Best_mimo_cond_2x2_opt.pth): fresh per-controller generation time [cuda]: mean 1.0083 ms (sd 0.1485 ms over 3 draws) original run's stored gen_ms (results/mimo/regret.npz, unrounded): 1.7533 ms existing per-plant oracle time (results/mimo/regret.npz, unrounded): 1440.9613 ms ratio (oracle / fresh-gen-1-ctrl) [cuda]: 1429.1x ratio (oracle / fresh-gen-K=8-ctrl) [cuda]: 178.6x (the K-controller figure is the fairer 'cost to actually run the amortized synthesis' number; see RESULTS.md ยง4.10's own ~600x-vs-~7900x distinction) ratio using the ORIGINAL run's own numbers (resolves the NIT: 1441/1.753 = 821.9x, not the rounded '2 ms -> 720x/822x' ambiguity) pendulum (models/Best_pendulum_cond.pth): fresh per-controller generation time [cuda]: mean 1.3702 ms (sd 0.0795 ms over 3 draws) original run's stored gen time (results/pendulum/eval.npz, unrounded): 2.0460 ms existing per-plant oracle time (results/pendulum/eval.npz, median of oracle_t, unrounded): 16089.6220 ms ratio (oracle / fresh-gen-1-ctrl) [cuda]: 11742.3x ratio (oracle / fresh-gen-K=8-ctrl) [cuda]: 1467.8x ratio using the ORIGINAL run's own numbers: 7863.8x saved -> results/stats/timing_cuda.{txt,npz}